Citations

Abstract Computer vision now enables deep learning models to link histopathologic features to molecular features in a quantitative manner not previously achievable by human pathologists. However, human interpretability is often limited. We trained a deep neural network to predict gene expression from digital whole slide images (WSIs) using 848 central nervous system (CNS) tumors with paired WSI and RNA-sequencing data. We used inferred RNA expression levels as a surrogate for protein expression of the CNS tumor lineage markers GFAP, OLIG2, and SSTR2, as well as the proliferation marker MKI67 (Ki67). We established thresholds for categorizing the inferred expression levels as positive or negative based on levels observed in cross-validation testing in select tumor types to approximate routine clinical interpretation of immunohistochemical (IHC) staining. We tested the sensitivity and specificity of our ‘virtual’ IHC on an independent, multi-institutional cohort of over 2,000 CNS tumors with objective diagnostic labels derived from DNA methylation-based tumor classification. As a dichotomous variable, inferred GFAP expression showed a sensitivity of 74% and a specificity of 99% as a glial marker in a cohort of gliomas and meningiomas. In the same cohort, OLIG2 showed a sensitivity of 75% and a specificity of 97% as a glial marker, while SSTR2 showed a sensitivity of 83% and a specificity of 97% as a marker of meningioma. In a cohort of ependymomas and gliomas, virtual IHC for OLIG2 showed a sensitivity of 75% and a specificity of 82% in distinguishing ependymomas from gliomas. Finally, in a cohort of gliomas, inferred expression of MKI67 demonstrated a trend consistent with what would be expected by actual IHC, with increasing MKI67 expression with increasing tumor grade. Our model provides the basis for a human interpretable and clinically applicable deep neural network to aid human pathologists in the diagnosis and grading of CNS tumors.

Abstract BACKGROUND Glioblastoma (GBM) is a molecularly and phenotypically heterogeneous tumor, with regional invasion patterns driving treatment resistance and universal recurrence. In infiltrative regions, GBM cells adopt at least two predominant invasive phenotypes: Neuronal-enriched (NEU), with synaptic and neurodevelopment programs, and Glycolytic/Plurimetabolic (GPM), with heightened immune and metabolic activity. These subtypes differ in therapeutic vulnerability and microenvironmental interactions; features not captured by conventional contrast-enhanced (T1+C) MRI due to minimal blood-brain barrier (BBB) disruption. Our group sought to develop predictive models to distinguish variability of NEU/GPM subtypes within infiltrative GBM, utilizing graph convolutional networks (GCN), advanced multiparametric MRI (mpMRI), and spatially-matched image-localized biopsies with transcriptomic profiling. METHODS We analyzed 18 mpMRI features —including conventional, diffusion-weighted, and Dynamic Susceptibility Contrast (DSC) perfusion MRI— of 101 image-localized biopsies from MRI non-enhancing infiltrative margins of 64 GBM patient tumors. Features were hierarchically clustered per modality; each cluster was represented as a node in a Neo4j graph. Edges connected nodes sharing overlapping samples. Nodes were annotated with NEU or GPM subtype proportion. GCNs were trained to classify subtype-enriched nodes using graph topology and node features. RESULTS GCNs achieved high predictive performance on held-out test data (AUC = 0.95 for NEU, 0.88 for GPM; 70:30 train-test split) with similar 5-fold cross-validation performance and generalized well to an independent cohort (AUC > 0.85, n = 45). Importantly, performance was robust without relying on T1+C imaging, enabling prediction beyond BBB-disrupted regions. GCNs outperformed traditional classifiers (logistic regression, random forest). CONCLUSION Our results demonstrate that GCN-modeling of mpMRI features enables non-invasive prediction of invasive GBM subtypes at the tumor margin. Unlike convolutional neural networks (CNN), which rely on spatial continuity, GCNs leverage relational structure between heterogeneous imaging features, making them ideal for modeling sparse, non-grid data. This may support non-invasive regional vulnerability detection, surgical targeting, and improved response assessment.

Abstract Molecular testing has become critical components in the diagnostic classification of the central nervous system (CNS) tumors. However, these methods require substantial resources, limiting accessibility for many patients. Recent advances in artificial intelligence (AI) and computer vision empower deep learning models to infer molecular features from histopathology images, which may often be more readily available than molecular testing, to classify CNS tumors. We trained pan-CNS tumor models to predict DNA methylation and gene expression from whole-slide images (WSIs). These molecular predictions were then used in a hierarchical machine learning framework, termed Neuropath-AI, to predict nine broad tumor families and 52 tumor types using a large diverse cohort of 5,715 histopathology samples, of which 2,988 had paired DNA methylation profiling and 848 had paired RNA-sequencing. We evaluated Neuropath-AI on another large independent multi-institutional cohort of 5,516 CNS tumors, for which Neuropath-AI predicted tumor class with an associated confidence score. Neuropath-AI designated 46% of test samples as predictable with high-confidence, which it then accuracy classified in 97% of samples. It made moderate-confidence (and above) predictions for 87% of test samples, for which it achieved a diagnostic accuracy of 80% for the top-1 prediction and 86% accuracy for a top-2 predictions. A comparison with human neuropathologists showed that Neuropath-AI is comparable at classifying tumors. A second test showed that human neuropathologists were more accurate when the Neuropath-AI classifier results were provided to them, testifying to the potential translational benefit of integrating AI classification tools into the pathologist’s workflow. Our model provides the basis for a clinically applicable deep learning assistant to improve human efficiency and accuracy of CNS tumor diagnosis. The model will be made publicly available and can be readily implemented to assist human pathologists in clinical workflows, in further expanded prospective studies.

Abstract There are no approved treatments for relapsed/refractory (R/R) PCNSL. We show promising efficacy in the salvage setting using emavusertib in combination with ibrutinib, which targets the BCR and TLR pathways. These 2 major pathways are associated with activation of NF-κB. Updated safety and efficacy data of this combination in R/R PCNSL from the ongoing Phase 1/2 TakeAim Lymphoma trial (NCT03328078) will be presented. As of 01 May 2025, 33 patients with R/R PCNSL were treated with emavusertib in combination with ibrutinib. Eligible patients receive emavusertib in combination with ibrutinib until tumor progression or unacceptable toxicity. Median age was 63 years (range 43-81); 16 patients were male. Median prior anti-cancer therapy was 2 (range 1-5); 26 (78.8%) patients had prior BTKi exposure. Median duration of emavusertib/ibrutinib treatment was 51 days (range 5-798). Grade ≥ 3 treatment-related adverse events (TRAEs) were reported in 11 (33.3%) patients; neutropenia in 5 (15.2%) patients, amylase increased, lipase increased, and leukopenia in 2 (6.1%) patients each. The majority of Grade ≥ 3 TRAEs were reversible and manageable. Among the 19 BTKi-experienced patients treated for ≥ 28 days, there were 5 CRs and 2 PRs, with an ORR of 36.8%. Three patients with CR were on treatment for 10+ mos, with 1 CR patient ongoing with DOT of 26.2+ mos. Among the 5 BTKi-naïve patients treated for ≥ 28 days, 1 achieved CR and 3 had PR, with ORR of 80%; the CR patient is ongoing with DOT of 18.1+ mos. The ORR in the combined population was 45.8% (6 CRs; 5 PRs). The combination of emavusertib and ibrutinib demonstrates promising clinical activity and an acceptable safety profile in patients with R/R PCNSL. This regimen appears capable of overcoming resistance to prior BTKi therapy. Patient enrollment in this trial continues.

Abstract PCNSL is a rare and aggressive NHL with no approved treatments for R/R patients, representing a critical unmet need. MyD88 mutations in ~70% of PCNSL patients drive IRAK4 activation, promoting NF-κB signaling, inflammation, and tumor progression. Emavusertib, a potent oral IRAK4 inhibitor, crosses the blood-brain barrier and shows preclinical synergy with BTKis, and has the ability to re-sensitize BTKi-resistant cell lines. This study evaluates the molecular and PK data associated with responses to emavusertib + ibrutinib combination therapy in R/R PCNSL patients. The safety, clinical activity, and potential biomarkers of emavusertib in R/R PCNSL are being investigated in the ongoing TakeAim Lymphoma trial (NCT03328078). Pre-dose and 1.5-hour post-dose plasma samples were collected on Day 1 of Cycles 3, 5, and 7. CSF samples were obtained via a lumbar puncture within 1.5 hrs of collection of the post-dose plasma PK sample on Cycle 3 Day 1. As of 06 December 2024, CSF concentration was available for 7 patients. The mean emavusertib concentration in CSF was 81.3 ng/ml (54.7-104.0) in patients receiving 100 mg emavusertib BID (n=4). In patients receiving 200 mg emavusertib BID (n=3), the mean emavusertib concentration in CSF was higher at 175.7 ng/ml (114.8-209.4), which is 2.2 × the mean value in patients who received 100 mg emavusertib BID (p-value = 0.02). MyD88 mutation status was available for 7 patients of which all had prior BTKi exposure. Among these patients, 6 had MyD88 mutations of which 4 had responded (3 CRs and 1 PR) to the combination with DOR up to 18.9 mos with data collection ongoing. Preliminary CNS PK data demonstrate that emavusertib concentration in CSF increases with increasing emavusertib dose. Patients with MyD88 mutations showed expected promising preliminary efficacy to emavusertib + ibrutinib combination and may overcome BTKi resistance. Enrollment is ongoing.

Abstract BACKGROUND Mirdametinib is the first FDA-approved MEK1/2 inhibitor for both adults and children ≥2 years with NF1 who have symptomatic PN not amenable to complete resection. We report results from the phase 2b ReNeu trial (NCT03962543), including the LTFU. METHODS In ReNeu, mirdametinib was administered in 3-weeks-on/1-week-off 28-day cycles. Following the 24-cycle treatment phase, patients could continue mirdametinib in an optional LTFU. We report results through June 12, 2024, approximately 9 months from the previous data cutoff. RESULTS Fifty-eight adults and 56 children received mirdametinib. Median (range) duration of treatment was 21.8 (0.4, 54.4) months in adults and 25.4 (1.6, 48.5) months in children. Among LTFU-eligible patients, 26/31 (84%) adults and 32/37 (86%) children entered LTFU. Confirmed objective response rate (ORR, treatment+LTFU phases) was 47% (27/58; 95%CI, 33%, 60%) in adults and 55% (31/56; 95%CI, 42%, 69%) in children. Among confirmed responders, 18 adults and 19 children achieved a deep response (best target PN volume reduction from baseline of >50%). Median (range) best percentage change from baseline in target PN volume was –41% (–90%, 13%) in adults and –43% (–98%, 48%) in children. Median duration of response (DoR) was not reached for either age cohort. Treatment-related adverse events (TRAEs) in ≥25% of patients were dermatitis acneiform, diarrhea, nausea, and vomiting in adults; and diarrhea, dermatitis acneiform, and paronychia in children. Grade ≥3 TRAEs occurred in 17% of adults and 25% of children. Dose interruptions, dose reductions, and discontinuations due to TRAEs occurred in 9%, 17%, and 22%, respectively, of adults, and 14%, 14%, and 9%, respectively, of children. CONCLUSIONS Improved confirmed ORR was observed with longer duration of mirdametinib, with additional deep responses achieved in both adults and children. Mirdametinib continued to be well tolerated with a manageable safety profile and no new safety signals in the LTFU.

Abstract BACKGROUND Mirdametinib is the first FDA-approved MEK1/2 inhibitor for both adults and children (≥2y) with symptomatic NF1-PN. In ReNeu, symptomatic improvement and PN volume reductions were observed at the earliest assessments (~91 and ~152 days, respectively) and were sustained. This analysis evaluated time to first onset and resolution of commonly reported TRAEs in the ReNeu trial. METHODS Data cutoff (DCO) was September 20, 2023. Median time to first onset and resolution (days) for common TRAEs (≥15%) were evaluated for prepubescents (≥2 to <12y, n=32), postpubescents (≥12 to <18y, n=24), younger adults (≥18 to ≤50y, n=48), and older adults (>50y, n=10). RESULTS Common TRAEs were mostly grades 1 to 2 (83%). In prepubescents, median day (range) of onset was 32 (14, 263) for vomiting, 71 (19, 386) for dermatitis acneiform, 123 (1, 663) for diarrhea, 124 (52, 775) for paronychia, and 127 (2, 663) for nausea. For postpubescents, median day (range) of onset was 8 (1, 438) for diarrhea, 8 (2, 574) for nausea, 12 (1, 252) for dermatitis acneiform, 28 (2, 354) for vomiting, and 110 (43, 245) for paronychia. In younger adults, median day (range) of onset was 5 (1, 352) for nausea, 8 (2, 15) for dermatitis acneiform, 9 (2, 73) for diarrhea, and 29 (1, 264) for vomiting. For older adults, median day (range) of onset was 4 (1, 286) for diarrhea, 9 (2, 15) for nausea, and 13 (2, 38) for dermatitis acneiform. Time to resolution was variable between TRAEs and among patients in all age groups; 90% resolved completely by DCO. Among all patients, 18 had a TRAE of asymptomatic ejection fraction decrease, with all but one resolving by DCO. CONCLUSIONS Overall, common TRAEs arose before treatment benefits and were generally resoluble by DCO. Proactively managing TRAEs may improve treatment experience and adherence.

Abstract Glioma stem cells (GSCs) are key drivers of glioblastoma (GBM) progression, therapy resistance, and recurrence, making them critical therapeutic targets. While the lineage-defining transcription factor OLIG2 is critical for GSC maintenance, the molecular co-regulators and downstream pathways it engages remain poorly defined. Here, we identify a previously unrecognized, functionally significant interaction between OLIG2 and STAT5B, uncovering a novel regulatory axis in GSCs. Using proximity ligation assay (PLA), we demonstrate that OLIG2 interacts with STAT5 in human and mouse GSC models, particularly in EGFRvIII-mutant cells. We find that OLIG2 preferentially partners with STAT5B and not STAT5A. Knockdown studies in GSCs reveal that OLIG2 and STAT5B co-regulate critical oncogenes BIRC5 and NUSAP1, both of which are required for sphere formation and migration capacity of EGFRvIII positive GSCs. In orthotopic xenograft models, knockdown of either STAT5B or OLIG2 significantly prolongs survival reduces expression of stemness and invasion markers, SOX2, VIMENTIN, and NUSAP1. In contrast, STAT5A knockdown fails to recapitulate these phenotypes, underscoring the role of OLIG2-STAT5B signaling in GSCs. These findings reveal a novel OLIG2-STAT5B transcriptional circuit that plays a role in GSC stemness maintenance, invasion, and tumor growth, particularly in EGFRvIII-driven gliomas. Targeting this axis may offer a new avenue for GBM treatment.

Abstract Glioblastoma (GBM) invasion is shaped by dynamic clonal evolution within the spatial and cellular architecture of the brain. To investigate this process, we integrated multiparametric MRI, whole-exome sequencing (WES), bulk RNA-seq and single-nucleus RNA-seq (snRNA-seq) from 69 IDH-wildtype GBM patients, using spatially resolved biopsies from contrast-enhancing (CE) core tumor and non-enhancing (NE) invasive regions. Phylogenetic analysis revealed distinct clonal architectures. CE-enriched clones carried truncal mutations in conanical drivers of cell cycle and RTK signaling (TP53, PTEN, EGFR, PDGFRA/B), and private alterations in DNA repair genes (CDH11, PAX7, RAD51B). In contrast, NE-enriched clones exhibited private mutations in MAPK and Hippo pathways components (STAG2, BRAF, MAP2K3, CTNNA1), suggesting region-specific evolutionary trajectories. Furthermore, snRNAseq cell-cell communication analysis uncovered enriched ligand-receptor signaling in NE regions between malignant cells and brain-resident populations, including PTN—PTPRZ1 (myeloid), NRG-ERBB4 and SLIT-ROBO (neuronal, astrocytic, and oligodendrocytic). Bulk RNAseq deconvolution confirmed increased representation of neurons and glial cells in biopsies harboring NE-specific clones. These findings define a unique population of GBM clones localized at the infiltrative margin, marked by private MAPK/Hippo mutations and enhanced intracellular communication with the surrounding brain microenvironment. The recurrent engagement of neural-glial signaling pathways suggests that these NE-associated clones co-opt normal cellular networks to support invasion, immune evasion, and persistence following surgical resection. This study establishes a mechanistic link between spatial clonal evolution and tumor-host communication at the GBM infiltrative region and provide a framework for targeting cell-cell signaling circuits that sustain invasive tumor ecosystems and escape therapeutic treatment.

Abstract IDH-wildtype glioblastoma (GBM) is the most malignant primary adult brain tumor, with a five-year survival rate below 5%. Despite maximal treatment, including surgical resection, radiotherapy, and temozolomide (TMZ), recurrence is universal, driven by profound intratumoral heterogeneity, diffuse invasion, and therapy resistance. A hallmark of IDH-wildtype GBM is frequent alteration of epidermal growth factor receptor (EGFR), including amplification (50%), the EGFRvIII deletion variant (25%), and extracellular missense mutations such as G598V and A289V (10–15%). These missense mutations increase ligand affinity and receptor activity, yet their downstream signaling consequences and role in shaping the tumor phenotypes remain poorly defined. Recent pathway-based subclassifications identified four subtypes in IDH-wildtype GBM. Notably, neuronal (NEU) and glycolytic/plurimetabolic (GPM) subtypes predominate the invasive tumor margins. NEU programs are enriched in synaptic and neurodifferentiation signals, while GPM programs are associated with metabolic adaptation, angiogenesis, and immune evasion. Here, we show EGFR-G598V and EGFR-A289V mutations differentially regulate these invasive subtypes identities. Transcriptional analysis reveals EGFR G598V upregulates metabolic and inflammatory pathways consistent with the GPM phenotype, while EGFR A289V activates genes involved in neuronal differentiation and synaptic function, aligning with the NEU subtype. Phosphoproteomics and kinase-substrate analysis further uncovered distinct kinases to be functionally activated/inhibited in EGFR A289V-expressing cells and G598V cells. Mechanistic studies revealed EGFR A289V induces sustained EGFR phosphorylation at Y992, activating PLCγ2 via the SRC-family kinase LCK. This cascade drives calcium release, CAMK1 activation, and STAT5 phosphorylation, defining a novel EGFR A289V–LCK/PLCγ2/CAMK1/STAT5 axis. In addition, Y1068 phosphorylation is also constitutively active in EGFR A289V, resulting in increased BRAF activation. In contrast, EGFR G598V activates MAPK6 and CAMK2 inducing AKT2 activation, reinforcing GPM-associated signaling. These findings uncover mutation-specific signaling networks that drive subtype fate in invasive GBM and highlight unique pathway(s) as tractable targets for subtype-directed therapy in EGFR-mutant GBM.